Spaces:
Sleeping
Sleeping
Commit History
-
Update README.md
Verified signatureb97e2f2 -
Use validated Build Small tag slugs (track/sponsor/achievement)
Codex4beb7c8 -
Add field notes (lessons learned) to README and FIELD_NOTES.md
Codexe9342cc -
Add demo video link to README
Codex85e7beb -
Warm boss card art in battle (GPU free post-draft) so boss plays show art instantly
Codex +1: Claude Opus 4.8Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>7e0483e -
Pre-bake backbone card art (instant, no shimmer) + stop warming face-down enemy hand
Codex +1: Claude Opus 4.8- 18 bundled JPEGs (6 backbone cards x 3 worlds); backbone_card is born with its art_uri, so the base Attack/Block/Draw cards never shimmer and cost no live SDXL. - Skip art generation for the enemy hand (rendered as card backs), freeing the art lane so visible player cards illustrate faster. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>a33d5b8 -
Speed up boss deck (parallel picks, ~22s->7.5s) and forge it from run start; deck-aware README
Codex +1: Claude Opus 4.8- draft_enemy_deck_for_ui: generate the 9 boss picks concurrently (1 card each, not deck-aware) instead of ~22s sequential; warm it from draft step 0 so it is ready before the battle handoff (no 'boss shuffles' wait). - make_pack: pack_size param. - README: emphasize deck-aware drafting (shapes packs to your build, dangles tempting off-archetype cards). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>372c910 -
Polish demo: boss reveal screen, portraits, varied packs, paced draft, title copy
Codex +1: Claude Opus 4.8- Boss reveal screen before rules: full splash art + "Your Enemy" + villain name and quote (The Hollow Warden / Specter-9 / The Necrolich); battle board shows face-cropped player vs boss portraits. - Draft pacing: minimum loading window per pick (uniform "forging" beat), prefetch all branches so any pick is ready, reveal first pack the moment it is ready. - Card variety: rotate per-slot effect focus per pack, and fall back to the slot focus (not always deal) so name-rejected packs still show varied effects. - Name filter bans the leaked word "primitive". - Title screen: tighter cohesive layout + model-stack pitch copy. - README: Build Small submission section, tags, model stack. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>39c3669 -
Fix card truncation: lead JSON with effects+name, request concise flavor/art, bump card tokens
Codex +1: Claude Opus 4.8Root cause of pool-name + junk-effect cards: the model wrote name/flavor/long art_prompt first and effects last, so max_tokens cut the JSON before it closed -> parse failed -> whole card discarded (pool name, salvaged stray words like 'energy'). Now the requested JSON leads with effects then name (they survive a cutoff), flavor/art_prompt are capped short, and card tokens raised to 320. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>3625c13 -
Trim prefetch to 1 pack, log model-pack failures, JSON-constrained cards, rotate fallback names
Codex +1: Claude Opus 4.8Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>b483ca7 -
Size Modal pools under the 10-GPU plan cap (card 3/4, art 3/4, boss 1/2)
Codex +1: Claude Opus 4.8min_containers=5 x3 = 15 exceeded the 10-GPU limit, so Modal throttled and starved art to 2 containers. Now 7 warm / 10 burst max: 3 warm card + 3 warm art cover a pack's parallel calls without tripping the cap. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>8f039b3 -
Salvage card effects from raw model text; min_containers=5 for demo
Codex +1: Claude Opus 4.8- generator: when MiniCPM emits malformed JSON or nests effects (no valid primitive list), scan the raw text for allowed primitive ids and use them, instead of collapsing every card to a default 'deal' (the monotone packs). - modal_app.py: min_containers 1 -> 5 so prefetch misses never hit a cold container during the demo (downgrade after). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>078e3e5 -
Speed: reveal first pack immediately, defer boss deck, more Modal containers
Codex +1: Claude Opus 4.8- ui.collect_ready_pack: stop withholding the first pack until speculative prefetch packs are ready (was generating ~12 cards before showing pack 1); reveal as soon as the pack itself is ready, prefetch in background. - ui.warm_enemy_deck: defer boss-deck forging to draft_step >= 2 so its ~9 card calls do not compete with the first visible pack (non-blocking handoff covers a late deck). - modal_app.py: max_containers 2 -> 8 (parallelize the burst); card default max_tokens 192 -> 128. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>c395e6f -
Hardcode Modal endpoint URLs in app_hf.py (public; env still overrides)
Codex +1: Claude Opus 4.8Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>0b3ffc4 -
Pin diffusers==0.31.0 on Modal art image (transformers 4.49 compat; avoids flux2/Qwen3 import)
Codex +1: Claude Opus 4.8Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>2fa3891 -
Add MODE switch: LOCAL (on-device, default) vs MODAL (GPU endpoints)
Codex +1: Claude Opus 4.8- clients.configure_mode(): MODE=LOCAL runs MiniCPM/Nemotron/SDXL in-process on your own hardware (the default on clone); MODE=MODAL routes cards/boss/art to Modal endpoints over HTTP. Per-var TABRAS_* overrides still win. - app.py __main__ -> LOCAL by default; app_hf.py -> MODAL on the Space. - Device-aware .to(best_device()) (cuda/mps/cpu) so LOCAL runs anywhere. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>8c897f9 -
Modal-backed Space: run models on Modal GPU endpoints, Space is a thin HTTP client
Codex +1: Claude Opus 4.8- modal_app.py: MiniCPM (cards) + Nemotron (boss) as OpenAI-compatible chat endpoints, SDXL-Turbo (art) returning a data URI; warm A10G containers + a cached HF volume. - art.py/clients.py: ModalArtClient + a 'modal' art backend. - app_hf.py: endpoint-based backends (card/boss over HTTP, art via modal); endpoint URLs come from Space variables. No GPU on the Space. - requirements: slim to gradio only; README -> cpu-basic, Modal-backed. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>5bdf0f6 -
ZeroGPU: move models to CUDA inside @gpu calls, not in the main process
Codex +1: Claude Opus 4.8ZeroGPU forbids real CUDA init outside an @gpu allocation. load() now keeps models on CPU (no .cuda(), no device_map=auto, no pipe.to('cuda')); the move to CUDA happens inside _minicpm_generate/_nemotron_generate/_run_art_pipe. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>c1939e1 -
ZeroGPU: read models from module globals, not pickled args
Codex +1: Claude Opus 4.8@spaces.GPU forks a worker and pickles args/returns across the process boundary; a trust_remote_code model (MiniCPMV) and a diffusers pipe can't be pickled. Move the @gpu boundary to thin module-level inference functions that read the model/tokenizer/pipe from module globals (inherited by the fork) and take/return only strings. complete()/create_art() set the global then call them; load() runs in the main process (no @gpu). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>f4b9db0 -
Pin transformers for MiniCPM-V-4; expose module-level demo for HF Spaces
Codex +1: Claude Opus 4.8- transformers==4.49.0: MiniCPM-V-4's trust_remote_code modeling targets the pre-meta-loader API (_tied_weights_keys); newer transformers crashes with 'MiniCPMV has no attribute all_tied_weights_keys' on load. - app_hf.py exposes a module-level `demo` and launches with css/head. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>0bd2bdc -
Integrate ZeroGPU Space path with card-cache, sampling, and diversity fixes
Codex +1: Claude Opus 4.8- ZeroGPU: zerogpu.py provides a `gpu` decorator (spaces.GPU on a Space, no-op locally) applied to art + transformers inference/load; README sets suggested_hardware: zero-a10g; requirements adds `spaces`. - clients.py: cache the Transformers MiniCPM load so the card client is not rebuilt (reloading the model) on every handler/timer tick — critical on ZeroGPU, where the @gpu .load would otherwise reallocate every 0.8s. - local_llm.py: MiniCPM card authoring now samples (temperature) instead of greedy decoding, alongside the @gpu decorators. - generator.py: per-slot primitive focus so a pack spans different effects, plus per-card fault tolerance (one failure no longer collapses the pack). - app.py: warm models on Play so the GPU loads during setup navigation. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>3780aa5 -
Update README
Vishnu Vennelakantiaee6435 -
Prepare Space deployment flow
b939433 -
Update public README
Codexfe0b867 -
finished base
a8afc36
-
Build Tabras card duel prototype
6bbf552
-
Update README.md
Vishnu VennelakantiUnverified signatured8e5b40 -
Initial commit
Vishnu VennelakantiUnverified signature5b5e1fd